Page 173 - Read Online

Wang et al. J. Mater. Inf. 2026, 6, 13 Page 17 of 19

and label quality, and data fusion strategy are investigated through experiments. The main findings are as
follows:

Physical synthetic image generation: A physically-based method is proposed for scratch segmentation on
free-form surfaces with limited data. Results show that this approach can significantly expand the coverage of
possible scratch image distributions and improve segmentation robustness compared to data-driven
methods.

Data quality matters. Higher similarity between physical synthetic data and real images improves
segmentation accuracy. However, the impact diminishes as the dataset size increases.

Key factors for integrating synthetic data. Transformer-based models are preferred to capture global scratch
features. Label quality and coverage also influence performance. Transfer learning can substantially improve
model generalization on real datasets.

For future work, scratches should be defined more flexibly to simulate complex scenarios, such as cross
scratches and shallow micro scratches. The method could also be extended to other defects, such as dents,
pits, and cracks. Developing a comprehensive defect model library covering typical aerospace alloy products
for synthetic image generation would facilitate the application of AVI systems in real aerospace production.
Additionally, simulating penetrating radiation for internal defects represents another important area for
exploration.

DECLARATIONS
Authors’ contributions
Made substantial contributions to the conception and design of the study and performed data acquisition
and interpretation: Wang, Y.; Wang, P.; Bai, Y.; Wang, W.; Zhu, M.
Made substantial contributions to literature analysis and manuscript writing: Wang, Y.; Wang, P.; Bai, Y.
Performed results analysis, discussion, and manuscript revision: Wang, Y.; Wang, W.; Bai, Y.; Zhu, M.; Luo, M.
Provided administrative, technical, and material support: Wang, W.; Zhu, M.; Luo, M.

Availability of data and materials
The data are available from the corresponding author upon reasonable request.

AI and AI-assisted tools statement
Not applicable.

Financial support and sponsorship
This work was supported by the National Natural Science Foundation of China (No. 52375516; No.
52005439) and the Natural Science Basic Research Program of Shaanxi (No. 2025JC-YBQN-649).

Conflicts of interest
All authors declared that there are no conflicts of interest.

Ethical approval and consent to participate
Not applicable.

Consent for publication
Not applicable.

Copyright
© The Author(s) 2026.